7 registered · 4 with at least one run on disk (4 reaching EVALUATION tier) · 3 never executed. A hypothesis whose falsifier has not been executed is never reported as supported, and is counted here as debt rather than omitted.
| id | claim | tier | m / n | runs | status |
|---|---|---|---|---|---|
| V1 "SSL beats handcrafted features" may be an artifact of leaky | Under simultaneously speaker-disjoint and demographically balanced splits, frozen SSL/health embeddings do not reliably beat eGeMAPS. | EVALUATION | m = 9 (3 encoders × 3 corpora, each vs eGeMAPS) · n = 10 → 0.001953 ≤ 0.05/9 = 0.005556 ✓ | v1-ssl-vs-handcrafted n=10 | CLAIM SUPPORTED |
| V2 Speaker-identity subspace ablation: how much of the "disease" | A large fraction of frozen-embedding disease discrimination lives in a low-rank speaker-identity subspace, and removing it collapses AUC further than a variance-matched random projection of the same rank does. | EVALUATION | m = 14 (7 ranks k ∈ {1,2,4,8,16,32,64} × 2 corpora) · n = 10 → 0.001953 ≤ 0.05/14 = 0.003571 ✓ | v2-shuffle-control n=3 v2-speaker-subspace n=10 | CLAIM SUPPORTED |
| V3 Where does health-specific pretraining actually cross general-purpose | Health-acoustic pretraining buys sample-efficiency, not ceiling: HeAR leads at small labelled-set size and is overtaken by Whisper-enc as n grows. | EVALUATION | m = 42 (2 contrasts {HeAR vs Whisper-enc, HeAR vs WavLM} × 7 sizes × 3 tasks) · n = 12 → 0.000488 ≤ 0.05/42 = 0.001190 ✓ | none | UNTESTED |
| V4 Does calibration degrade *with* distribution shift, or independently o | Calibration error does not track AUC collapse under corpus shift, so model confidence is unusable as an out-of-domain detector. | EVALUATION | m = 6 (3 corpus pairings × {raw, temperature-scaled}) · n = 10 → 0.001953 ≤ 0.05/6 = 0.008333 ✓ | none | UNTESTED |
| V5 Zero-shot audio-LLMs as a leakage-free reference point | The gap between a supervised probe and a zero-shot audio-LLM is a cheap, model-free estimate of how much leakage a protocol admits, because a zero-shot model has no training split and therefore cannot leak speaker identity, demographics, or preprocessing statistics. | EVALUATION | screening n ≤ 3 · promotion target m = 3 (3 datasets), n = 10 → 0.001953 ≤ 0.016667 ✓ (n = 8 would also satisfy m = 3) | none | UNTESTED |
| V6 Preprocessing-fit leakage is corpus-specific and cannot be assumed sma | The inflation from fitting the scaler/normaliser before splitting varies by an order of magnitude across corpora, so "we scaled before splitting but it's a small effect" is not defensible on an unmeasured corpus. | EVALUATION | m = 6 (3 corpora × 2 representations) · n = 10 → 0.001953 ≤ 0.008333 ✓ | v6-preprocessing-leakage n=10 | PARTIAL |
| V7 The Clever-Hans silence shortcut beyond the Pitt corpus | Non-phonatory signal alone (silence/pause structure, recording-level duration and intensity statistics) achieves a substantial fraction of headline performance on voice-health corpora outside the one where the effect was first documented. | EVALUATION | m = 6 (3 corpora × 2 shortcut feature sets) · n = 10 → 0.001953 ≤ 0.008333 ✓ | v7-silence-shortcut n=10 | PARTIAL |
Click a column header to sort; type to filter. Status is read from the ## Summary table of IDEA_TABLE.md; runs are the artifacts declared in scripts/build_common.py, so a completed run that no generator knows about fails the build rather than disappearing from this column.